What Is Astra? OpenAI's GPT-6 Flagship Model, Explained

Astra 101: OpenAI's Next Major Model That Solved 10 Open Math Problems for $2,000

Astra is OpenAI's flagship model family, first surfaced publicly on August 1, 2026, when an internal build was credited with producing 10 advances in mathematics and theoretical computer science for roughly $2,000 in total inference cost. OpenAI officially launched it as GPT-6 Astra on September 3, 2026. Astra is now available in ChatGPT Work, Codex, and the API, and the rollout reached all Plus and Business users by September 4. Its public API starts at $10 per million input tokens and $50 per million output tokens.

Updated 2026-10-06: GPT-6 Astra is live across OpenAI's core work and developer products. OpenAI has also increased its default generation speed by about 50% across subscriptions and partner products that use Sign in with ChatGPT.

Key Takeaways

• OpenAI officially launched GPT-6 Astra on September 3, 2026.
• An internal version produced 10 advances in mathematics and theoretical computer science, including work associated with quantum complexity.
• Total inference cost for those 10 results was about $2,000 at Sol API prices, per Greg Brockman.
• OpenAI positions Astra as its most capable model and as a larger, more expensive model class than Sol.
• GPT-6 Astra is the official public product name; Astra also refers to the broader model family.
• Published benchmark results include 99.9% on ARC-AGI-3 and 100% on ExploitBench.
• Astra is designed for computer use, software engineering, visual understanding, science, long-running agentic work, and difficult reasoning.
• Astra is available in ChatGPT Work, Codex, the API, and paid ChatGPT plans, including Plus, Pro, and Business.
• The API model ID is gpt-6-astra, with pricing starting at $10 per million input tokens and $50 per million output tokens.
• OpenAI says Astra reached the Critical cybersecurity capability threshold under its Preparedness Framework.
• Astra has a 1,050,000-token context window, with up to 922,000 input tokens and 128,000 output tokens.
• OpenAI has not formally designated Astra as AGI or ASI, and independent testing continues to show uneven performance across tasks.

What Is Astra?

Astra is the name of OpenAI's flagship model family. The name entered public view on August 1, 2026, when Greg Brockman described an internal version producing 10 significant advances in mathematics and theoretical computer science. OpenAI then officially launched the model as GPT-6 Astra on September 3.

The launch settled the model's status and branding. Astra is no longer an internal-only project or a rumored GPT-6 candidate: GPT-6 Astra is the public product name. OpenAI positions it as its most capable model for complex work, with a particular focus on computer use, browsing, professional work, software engineering, cybersecurity, science, visual understanding, and long-running agentic tasks.

The original math result remains the clearest reason Astra attracted attention. An internal version was credited with producing 10 results at an inference cost of about $2,000 at Sol API prices. The work spans mathematics and theoretical computer science, including quantum complexity, and reports connected some of the results to formal verification artifacts such as Lean certificates.

Astra is now broadly usable through OpenAI products. It is available in ChatGPT Work, Codex, and the API. OpenAI completed the rollout to all Plus and Business users on September 4, while Pro users also have access. Enterprise and cloud deployment channels were included in the broader launch program, although product availability and capacity can still vary by channel.

Astra at a Glance

Attribute	Value
Developer	OpenAI
Official name	GPT-6 Astra
Type	Flagship frontier reasoning and computer-use model family
Related models	Sol, Terra, Luna
Positioning	OpenAI's most capable model; a larger and more expensive model class than Sol
First public surfacing	August 1, 2026, through the math-results announcement
Launch date	September 3, 2026
Release status	Launched and publicly available
GPT version branding	Officially GPT-6 Astra; Astra also names the model family
Modalities	Text and image input; text output; no native audio or video input
Context window	1,050,000 tokens
Maximum input	922,000 tokens
Maximum output	128,000 tokens
Knowledge cutoff	April 30, 2026
API model ID	gpt-6-astra
Pricing	Starts at $10 per 1M input tokens and $50 per 1M output tokens
API availability	Live
ChatGPT availability	Available in ChatGPT Work and to Plus, Pro, and Business users
Codex availability	Available
Azure and AWS	Launch and customer availability have been announced or reported; specific access can vary
License	Closed and proprietary; no official open-weights release
Primary demonstrated capabilities	Computer use, browsing, coding, visual understanding, long-running agents, science, cybersecurity, and formal reasoning

How Astra Works and What Makes It Different

OpenAI has not published a complete architecture description. Industry reporting and community analysis describe a technique called recurrent depth, in which information is processed through model layers multiple times. The technique is reported to improve performance and token efficiency while making some internal reasoning less visible to external chain-of-thought monitoring.

That tradeoff matters for a model with strong agentic and cybersecurity capabilities. OpenAI launched Astra with additional safeguards, including monitoring intended to detect and contain potentially unauthorized activity. The company also limits access to Astra's most advanced cybersecurity capabilities more tightly than access to the general model.

OpenAI said Astra was created with its largest-scale training run to date, using more than 100,000 GPUs and extensive assistance from previous generations of AI models. The total parameter count, however, remains undisclosed. Pre-launch claims that Astra was a 10-trillion-parameter model were speculative and should not be treated as a confirmed specification.

The economics are now clearer than they were before launch. Brockman's original post described "ten significant advances in mathematics and theoretical computer science… solved using an internal version of Astra, our next major model, for a total cost of about $2000 at Sol API prices." That works out to roughly $200 per result. The released API separately starts at $10 per million input tokens and $50 per million output tokens.

Astra is also a model family rather than one immutable checkpoint. OpenAI has compared the naming structure to Sol, where multiple versions can exist under the same family name. Astra occupies a larger and more expensive class, while Sol, Terra, and Luna cover lower-cost or more routine work.

The most repeated system-level description is an orchestrator coordinating agents and tools. The released model does manage subagents and long-running workflows, but OpenAI has not published a complete account of how many agents it can coordinate, which models sit beneath it in each product, or how the orchestration layer is implemented.

What You Can Do With Astra

Astra is available for real workloads rather than being limited to internal demonstrations. Its observable capabilities include:

• Computer use: OpenAI describes Astra as capable of working through the same applications people use every day, including software without an API. Published results include 92.7% on ScreenSpot-Pro and 72.6% on OSWorld 2.0, with lower simulated latency than GPT-5.6 Sol.
• Formal mathematics and theoretical computer science: An internal version was credited with 10 advances spanning areas such as group theory, high-dimensional geometry, and quantum complexity. Reports linked the work to formal verification through Lean certificates, although some individual proofs and their provenance have drawn scrutiny.
• Software engineering and coding: OpenAI positions Astra as a major model for software engineering. Demonstrations show it creating complex software, generating games, working with SVGs and shaders, developing mobile software, and building or modifying 3D projects.
• Long-running agentic work: Astra is designed to continue working for days or weeks, remember corrections, coordinate subagents, and recover from mistakes. OpenAI has discussed an eventual version that could run continuously, but a universally available “runs forever” mode remains a future goal.
• Professional work: Astra can create and edit presentations, spreadsheets, documents, financial models, and branded materials. OpenAI says it can work inside existing business applications rather than requiring every workflow to be rebuilt around an API.
• Visual and 3D work: Public examples include Blender scenes, architectural visualizations, editable 3D objects, interactive environments, browser games, and programmatic animation. Results are often strongest when Astra is given references or an existing design to improve.
• Cybersecurity: OpenAI says Astra is its first model to reach the Critical cybersecurity capability threshold under its Preparedness Framework. It scored 100% on ExploitBench, and an internal evaluation included two newly discovered zero-day vulnerabilities. The most advanced cyber capabilities remain subject to tighter access controls.
• Science and research: Astra has been used for scientific-paper checking, medical education, drug visualization, technical forecasting, and long-horizon research tasks.

Astra's modalities also have an important limitation. It accepts text and image input and produces text output, but it does not support video as a native input modality. It can still analyze videos indirectly by using tools and extracting information through supported modalities, but that is not the same as native video understanding.

How Astra Compares

Astra's launch results make comparisons more concrete, but they do not produce one simple leaderboard. OpenAI and benchmark accounts report strong results across mathematics, computer use, cybersecurity, and professional work, while independent testing shows that its advantage depends heavily on the task and harness.

Astra's clearest first-party comparison is with GPT-5.6 Sol in cybersecurity and computer use. OpenAI describes Astra as both more capable and more token-efficient at vulnerability identification and exploit development. It achieved 100% on ExploitBench, while an internal benchmark using 20 more recently disclosed high-severity vulnerabilities found much higher arbitrary code-execution rates than Sol.

The ARC-AGI-3 result has also matured beyond the conflicting launch-day reports. A published benchmark account now gives Astra a 99.9% score. In one documented game, Astra invented compact algebraic notation for tracking state and cleared all eight levels in 242 moves without a reset. That is a strong result, although it remains one benchmark rather than proof of general intelligence.

Coding comparisons are more mixed. Community results report 74.1% on DeepSWE v1.1 and 57.9% on Terminal-Bench 4.0, with strong token efficiency. Some users prefer Astra for computer use, research, test design, data analysis, and fast execution, while others still favor Fable or Opus for large codebases, sustained implementation, or polished first-pass frontend design.

Dimension	GPT-6 Astra (OpenAI)	Fable 5.1 (Anthropic)
Status	Publicly launched and available	Publicly available
Reported strengths	Computer use, mathematics, science, cybersecurity, research, data analysis, visual understanding, and long-running agents	Large-codebase work and complex coding execution in several community comparisons
Benchmark signals	99.9% ARC-AGI-3; 100% ExploitBench; 74.1% DeepSWE v1.1 reported	Results vary by benchmark and harness
API price signal	Starts at $10 per 1M input tokens and $50 per 1M output tokens	Task-level comparisons vary
Access	ChatGPT Work, Codex, API, Plus, Pro, and Business	Existing public access
Evidence status	First-party disclosures, published benchmarks, and broad public use	Comparative judgments remain task-dependent

Independent tests also identify limitations. Astra scored 14% on the difficult MazeBench spatial-reasoning benchmark despite leading the tested models, and its Elo change was negative in one 200-game continuous-learning chess experiment. Users have also reported overworking tasks, consuming substantial quota, stopping early in some configurations, and needing follow-up prompts on complex builds.

The reasonable conclusion remains narrower than “Astra wins everything.” Astra is a major step forward in computer use, professional work, science, and cybersecurity. It can be exceptionally effective when paired with the right harness and task definition, but it still has a jagged capability profile and does not dominate every coding, design, learning, or long-horizon evaluation.

Availability: How to Access Astra

GPT-6 Astra launched on September 3, 2026, and the initial staged rollout has progressed into public availability:

• ChatGPT Work: Astra is available for professional workflows that combine reasoning, files, computer use, and business applications.
• Codex: Astra is available for software engineering, computer-use tasks, subagent coordination, and long-running projects.
• ChatGPT plans: The rollout reached all Plus and Business users on September 4. Pro users also have access, including the GPT-6 Pro experience associated with Astra.
• API: The gpt-6-astra model is live. Pricing starts at $10 per million input tokens and $50 per million output tokens.
• Partner products: Improvements also apply to products using Sign in with ChatGPT, including OpenCode, Pi, Amp, and Devin.
• Azure and AWS: Astra availability through these cloud channels has been announced or reported, although precise regional access and capacity can vary.
• Cybersecurity tools: Astra's most advanced cybersecurity capabilities remain more restricted than general model access because the model meets OpenAI's Critical threshold under its Preparedness Framework.

The temporary banked-reset program applied while paid users were waiting for rollout access; it is not Astra's permanent access model. OpenAI has since adjusted plan allowances, efficiency, and generation speed, including an approximately 50% default speed increase announced in October.

Developers can now use Astra alongside currently available frontier models such as GPT-5.6, choosing between maximum capability and lower-cost routine work according to the workload.

What We Don't Know Yet

The launch resolved Astra's existence, date, public name, context window, API availability, and base pricing. The remaining questions are narrower:

• Complete architecture and parameter count. Recurrent depth has been reported, but OpenAI has not published a full architecture description or confirmed the total parameter count.
• Long-term reliability. Public use has produced impressive results alongside failures in spatial reasoning, continuous learning, large-codebase work, and long-running execution. Broader independent replication is still needed.
• Multi-agent implementation. Astra can coordinate subagents, but the production architecture, agent-count limits, routing rules, and division of labor with Sol, Terra, and Luna remain partly undisclosed.
• Persistent-agent limits. Astra can run for extended periods, but the exact practical limits of memory, continuous operation, recovery, and unattended execution vary by product and harness.
• Current quotas and capacity. OpenAI has changed subscription allowances and usage accounting since launch, and heavy Astra workloads can still consume plan capacity quickly.
• Safety generalization. OpenAI added monitoring and access controls after designating Astra Critical for cybersecurity, but the long-term effectiveness of those safeguards remains an open research question.
• AGI status. OpenAI has not formally designated Astra as AGI or ASI. Whether it satisfies a particular AGI definition remains a matter for independent evaluation and debate.

For deeper context on how a GPT-6-class model could look, see coverage of what a GPT-6-class model could look like.

Frequently Asked Questions

What is Astra?

Astra, officially GPT-6 Astra, is OpenAI's flagship model family launched on September 3, 2026. It first surfaced publicly on August 1 through an internal version credited with 10 advances in mathematics and theoretical computer science. The released model focuses on computer use, reasoning, coding, visual understanding, science, and long-running agentic work.

Is Astra the same as GPT-6?

Yes. OpenAI officially launched the model as GPT-6 Astra. “Astra” also refers to the broader model family, so both Astra and GPT-6 Astra are used, but the public product name is no longer an unresolved rumor.

How much does Astra cost to run?

The public API starts at $10 per million input tokens and $50 per million output tokens. Separately, the internal run behind the 10 math and theoretical-computer-science results cost about $2,000 in inference at Sol API prices; that research-run total is not the API price.

When was Astra released?

OpenAI launched GPT-6 Astra on September 3, 2026. It is available in ChatGPT Work, Codex, and the API, and the rollout reached all Plus and Business users by September 4; Pro access is also available.

Is Astra AGI or ASI?

No official OpenAI designation says Astra is AGI or ASI. It has unusually strong computer-use, science, mathematics, coding, and cybersecurity results, but independent tests show a jagged profile, and whether it meets an AGI definition remains debated.

What is Astra good at?

Astra is strongest in computer use, browsing, professional work, software engineering, cybersecurity, science, visual and 3D workflows, and long-running agents. It scored 100% on ExploitBench and 99.9% on ARC-AGI-3, and an internal version was credited with 10 advances in mathematics and theoretical computer science. Its most advanced cyber capabilities remain more restricted.

How is Astra different from Sol, Terra, and Luna?

OpenAI positions Astra as a larger, more expensive model class than Sol. Sol, Terra, and Luna are related models aimed at lower-cost or more routine work, while Astra is used for the hardest reasoning, computer-use, planning, and orchestration tasks. OpenAI has not published the complete internal division of labor.

What to Watch Next

The next phase is about how Astra performs after its launch-day novelty fades. Independent tests already show a wide spread: benchmark-leading computer use, cybersecurity, and abstract reasoning coexist with weaker continuous learning, variable large-codebase performance, and occasional overexecution. Longer-term evaluations should reveal which gains come from the model itself and which depend on Codex, computer-use tools, subagent routing, or specialized harnesses.

Capacity and economics are another defining signal. Demand for Astra was high enough that OpenAI warned it might pause new Pro subscriptions, and plan allowances have changed since launch. OpenAI has improved generation speed and long-tail usage efficiency, but the practical cost of sustained Astra work remains important for both API customers and subscribers.

The cybersecurity rollout also remains consequential. OpenAI classified Astra as Critical under its Preparedness Framework and restricted the most advanced cyber capabilities while deploying additional monitoring. How those controls perform in production will matter as much as the raw ExploitBench score.

Finally, practical reliability will determine whether Astra becomes a durable default for complex work. Its strongest demonstrations span computer use, mathematics, coding, visual work, games, medicine, and science, but the most valuable evidence will come from repeatable results on real projects rather than isolated showcase outputs.

Update — 2026-08-22

Since publication, the Astra rumor cycle has escalated sharply, though the model remains unreleased with no confirmed date as of August 22, 2026. OpenAI staff and prominent leakers began openly "vagueposting" about an imminent release in mid-August (@AndrewCurran_, @kimmonismus), with some accounts claiming a launch "this week" or by month's end (@mark_k, @LuminaXspace). Those claims are contested: @DanDr1s said August was off the table, @kimmonismus called it still "weeks" away on August 20, and @pankajkumar_dev pointed to a first- or mid-September window. None are official.

The dominant explanation for the slippage is safety and cybersecurity. OpenAI published a cyber-capabilities post dated August 7, 2026, and multiple accounts tie Astra to a reported two-week pause on reinforcement-learning training for its latest deployment models, with the largest planned frontier RL run said to remain on hold (@kimmonismus, @LuminaBench). @mark_k cited Axios reporting that access may be delayed or restricted over cybersecurity concerns. All of these capability and delay claims come from third-party accounts and remain unverified.

A few structural details have also surfaced, all unconfirmed. Several leakers now describe "Doug" as the codename for Astra's pretraining run (with "Mewfour" as an internal checkpoint), rather than a separate model, and repeat a roughly 10T-parameter figure (@ChrisGPT, @LuminaBench). The most repeated architectural claim is that Astra will act as an agent-swarm orchestrator sitting above Sol, Terra, and Luna (@mark_k, @haider1) — consistent with, but not confirmation of, the "parent model" framing already discussed above. No pricing, benchmarks, or public availability have appeared as of August 22, 2026.

Update — 2026-09-06

On September 5, Code Arena placed Astra first in WebDev with 1,797 points, ahead of Claude Fable 5.1 at 1,762 and Claude Opus 5 at 1,688. The post cited a cost of $40 per million tokens. This is a community leaderboard rather than a standardized OpenAI evaluation, so it adds a counterpoint to earlier coding comparisons without settling Astra's broader coding advantage.

Separately, @kimmonismus claimed that Astra's pre-release use gave OpenAI a major competitive edge and increased internal productivity enough to move some plans forward by six months, from mid-next year to DevDay. That is an unverified account of internal impact, not an independently measured result.

Building similar advanced mathematical reasoning workflows today? On kie.ai you can try GPT-6 Astra, GPT-5.6, and Claude Opus 5.